An adaptive two-stage finite element model updating method fusing Bayesian neural networks and trust region resampling

Abstract This paper addresses the challenge of obtaining accurate finite element model updating (FEMU) results for complex flexible aircraft structures without relying on prohibitively large numbers of direct finite element evaluations. An adaptive two-stage framework is proposed by integrating Bayesian regularisation neural networks (BR-ANN) with trust-region resampling. In the global stage, a coarse-grained surrogate model is constructed via Latin hypercube sampling (LHS) and Bayesian inference, enabling the non-dominated sorting genetic algorithm II (NSGA-II) to identify the Pareto-front region containing promising candidate solutions. Subsequently, a trust-region strategy performs high-fidelity resampling within the local neighbourhood of the knee-point solution, establishing a refined surrogate model that reduces local surrogate bias and improves parameter convergence. Validation using the GARTEUR SM-AG19 benchmark aircraft demonstrates that the proposed method effectively mitigates the smoothing effect inherent in conventional global surrogate models while maintaining computational efficiency. The average modal frequency error was reduced from 4.56% to 2.39%, and the method supports physically interpretable updating of material and equivalent beam parameters, providing a reliable baseline for subsequent aeroelastic predictions.